Crop and weed stem classification with convolutional long short-term memory networks
LE3 .A278 2020
Master of Science
Mathematics and Statistics
Mathematics & Statistics
This study applies deep learning neural networks to the problem of identifying crop and weed stem origin points in images. An autonomous weeding robot drives through a field, taking overhead images as it drives over the crop row. A sample of images, with weed and crop labels added by an operator, is used to train a neural network. The trained neural network determines the locations of weed stem emergence points, then a mechanical picking arm travels to the location of the weed and pulls it from the ground. The sequential nature of the image data over time raises the question of whether learning sequential features in addition to local features of the data could improve weed classification. This thesis offers an investigation into the utility of convolutional long short-term memory layers for classifying crop and weed stem origins from a sequence of two images and introduces a novel weed/crop confusion matrix for evaluating model performance.
The author grants permission to the University Librarian at Acadia University to reproduce, loan or distribute copies of my thesis in microform, paper or electronic formats on a non-profit basis. The author retains the copyright of the thesis.